We consider a generic two-hop transmission setup. Explicitly, a source signal is transmitted over an imperfect channel, yielding a noisy observation. This signal shall then be compressed at a relay node before getting transmitted further over an error-prone and rate-limited channel to the sink, where the source signal is decoded/reconstructed. In [1], [2], we presented a data-driven Information Bottleneck-based quantization scheme called Deep FAVIB. For that, we derived a tractable variational lower-bound of the original objective functional that could be optimized using samples and utilized Deep Neural Networks (DNNs) to realize both the quantizer/encoder at the relay node and the decoder at the sink. Based on this work, we now provide further investigations, showcasing the excellent generalization capabilities of Deep FAVIB by several Symbol-Error-Rate (SER) simulation results. Specifically, we apply the pretrained Deep FAVIB to different environments that have not been present in the training, and show that yet, it yields promising results. This gives clear evidence to the fact that Deep FAVIB can be considered as a practically efficient scheme to be utilized, especially, when dealing with the highly dynamic and challenging environments.